Date: Wed, 11 Jul 2001 02:45:06 +0200
From: Andrea Arcangeli <firstname.lastname@example.org>
Diff between 2.4.7pre3aa1 and 2.4.7pre5aa1:
Only in 2.4.7pre3aa1: 00_async-io-unlock-race-1
Only in 2.4.7pre3aa1: 00_cpus_allowed-1
Merged in mainline.
Only in 2.4.7pre3aa1: 00_bh-async-1
Only in 2.4.7pre5aa1: 00_bh-async-2
Rediffed against pre5 due trivial rejects.
Only in 2.4.7pre5aa1: 00_create_bounces-sleeps-1
Kill KM_BOUNCE_WRITE and use kmap instead of kmap_atomic+cli in the
write-bounce case since create_bounces is allowed to sleep.
Only in 2.4.7pre5aa1: 00_drop___unlock_buffer-1
Rename __unlock_buffer to unlock_buffer. I did it before knowing Linus
did it too sorry.
Only in 2.4.7pre5aa1: 00_drop_async-io-get_bh-1
Drop the get_bh/atomic_inc around the async I/O, for any bh that uses
set_buffer_async_io the page will be locked for the whole duration of
the I/O (that's why you use set_buffer_async_io in first place) so the
bh doesn't need the additional b_count protection during the I/O and we
this way we can save some cycle.
Only in 2.4.7pre5aa1: 00_drop_end_buffer_write-1
Drop end_buffer_write and replace with the sync_io callback to avoid
wasting icache. Here too I should have waited for pre6 but I just
Only in 2.4.7pre5aa1: 00_drop_write_unlocked-get_bh-1
Micro-optimize a bit in write_unlocked_buffers: the get_bh is needed
only if the buffer was dirty and we are going to do the I/O on it (all
the rest is protected by the lru_list_lock).
Only in 2.4.7pre5aa1: 00_kiobuf-backout-get_bh-1
Remove the get_bh/put_bh around the kiobuf driven I/O, the notifion of
the refernece count and of I/O in flight belongs to the kiobuf not to
the bh, the bh are just an array of ram in the kiobuf, and this way we
save some cycle.
Only in 2.4.7pre3aa1: 00_ksoftirqd-7_ppc-2
Only in 2.4.7pre3aa1: 00_ksoftirqd-8
ksoftirqd is in mainline. All ports needs to synchronize now.
Alpha and x86 are uptodate in pre5. I uptodate uml in my tree
to get it right too.
For port maintainers: in short you need to do the:
check only when returning from irqs that can post softirq events (for
example IPI never post softirq events currently, theoretically you
could with the smp_call but nobody uses it that way at the moment). For
example x86 and alpha do the check in C in the last few lines of C of
the irq handler.
You _don't_ need to check for pending softirq when returning
from kernel to userspace any longer, that case is now handled with
ksoftirqd, this gives higher performance in the ret-to-userspace and
reschedule fast paths (that is the change made by Linus).
As usual ksoftirqd keeps handling the overflow case.
Only in 2.4.7pre5aa1: 00_linus-brelse-fix-1
This is a patch from Linus to fix a race condition he spotted in brelse
that could trigger on architectures where the atomic_t changes don't
imply SMP memory barriers at the CPU level.
Only in 2.4.7pre3aa1: 00_meminfo-wraparound-1
Only in 2.4.7pre5aa1: 00_meminfo-wraparound-2
Rediffed to fixup trivial rejects.
Only in 2.4.7pre3aa1: 00_msync-fb0-1
Merged in mainline.
Only in 2.4.7pre5aa1: 00_o_direct-10
Only in 2.4.7pre3aa1: 00_o_direct-9
Fixed O_SYNC|O_DIRECT, when O_SYNC was used in combination with
O_DIRECT it was returning a short write even if all the I/O was
reaching the disk succesfully and the f_pos was updated correctly, (in
my regression testing of O_SYNC|O_DIRECT I obviously forgotten to check
for short writes, I was only checking for errors and it was working
like a charm). Thanks to Peter for spotting and reporting this.
Fixed also another bug that could forget to wait for I/O completion
of async I/O for O_SYNC/fsync/fdatasync.
Only in 2.4.7pre3aa1: 51_uml-ac-to-aa-1.bz2
Only in 2.4.7pre5aa1: 51_uml-ac-to-aa-2.bz2
Updated to pre5 ksoftirqd logic.
A fixed O_DIRECT patch against vanilla 2.4.7pre5 is here: